Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Diffusion-Classifier Synergy: Reward-Aligned Learning via Mutual Boosting Loop for FSCIL
Authors: Ruitao Wu, Yifan Zhao, Guangyao Chen, Jia Li
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments |
| Researcher Affiliation | Academia | Ruitao Wu1,2 Yifan Zhao1 Guangyao Chen3 State Key Laboratory of Virtual Reality Technology and Systems, SCSE & QRI, Beihang University 2Zhongguancun Academy 3Peking University EMAIL |
| Pseudocode | Yes | Algorithm 1 FSCIL with DCS |
| Open Source Code | No | We commit to publicly releasing the source code immediately upon acceptance of our paper. |
| Open Datasets | Yes | Following the benchmark settings of previous methods, we conducted experiments on three datasets, i.e., mini Image Net [55, 45], CUB-200 [57], and CIFAR-100 [22]. |
| Dataset Splits | Yes | The division of the datasets aligns with existing methods. Specifically, the CIFAR-100 and mini Image Net datasets are partitioned into a base session containing 60 classes and incremental sessions containing 40 classes, with each session being an 8-way 5-shot few-shot classification task. The CUB-200 dataset is divided into the base session containing 100 classes and incremental sessions containing 40 classes, with each session being a 10-way 5-shot task. |
| Hardware Specification | Yes | All experiments were performed under Ubuntu 20.04.4 LTS operating system with NVIDIA Ge Force RTX 4090 GPU. |
| Software Dependencies | Yes | The experimental code is written in Python 3.8.19, and the Py Torch (version 1.13.0+cu117) is used for the deep learning framework. The source code of the diffusion model used in the experiments is from the open source library diffusers [1] (version 0.32.2). |
| Experiment Setup | Yes | In RPAMMD, α = 1, β = 2, and k = 0 (since this term is constant). In RRC, Tbase is set to 2.0, and Tscale is set to 1.0. In RCSCA, γ is set to 1.0, and only the top-3 old classes yt most similar to yc are considered. Training Details For the optimizer, SGD is used on all datasets with momentum of 0.9 and weight decay of 0.0005. In addition, we use the cosine annealing strategy to dynamically adjust the learning rate during training. For the CIFAR-100 dataset, the initial learning rate is 0.1 with 50 epochs for the base session, and 0.01 with 5 epochs for the incremental session. For the mini Image Net dataset, the initial learning rate is 0.1 with 120 epochs for the base session, and 0.05 with 30 epochs for the incremental session. For the CUB-200 dataset, the initial learning rate is 0.002 with 120 epochs for the base session, and 0.0005 with 10 epochs for the incremental session. Following [15, 68, 27], we employ Res Net-18 as the backbone for CUB200, and Res Net-12 for mini Image Net and CIFAR100. |